Unbiased Teacher v2: Semi-supervised Object Detection for Anchor-free and Anchor-based Detectors



CVPR 2022


Right: Anchor-free detectors are rapidly developed recently, while adapting the pseudo-labeling method on the anchor-free models results in less improvements compared with the anchor-based detectors.
Left: Illustration of Listen2Student mechanism, which explicitly compares the prediction uncertainties between the Teacher and the Student and selects these instances where the teacher has lower uncertainty than the student.




Motivation

Semi-supervised Object Detection aims to leverage unlabeled images to improve the supervised object detector. Despite the rapid development and significant improvement, there are still two remaining issues that are left untackled:
(1) There is no prior SS-OD work on anchor-free detectors
(2) Prior works are ineffective in pseudo-labeling on the bounding box regression.



Findings

llustration of Centerness bias issue. Left: Selecting pseudo-boxes based on box scores leads to worse results in semi-supervised learning compared with selecting based on classification scores. Center: Box scores of the anchor-free detectors are defined as the multiplication of the centerness scores and the classification scores, and we find that Right: box scores of pseudo-boxes is dominated by the centerness scores, which are unreliable in semi-supervised setting.

llustration of Unreliable label assignments. Left: Existing techniques for improving fully-supervised anchor-free detectors such as Center Sampling and localization-based classification labels are less robust to the localization noise (e.g., box center shifted) in pseudo-boxes, and the pixel-wise recall and precision of these two techniques are lower than the standard label assignment. Thus, our empirical evaluation shows that the standard label assignment leads to a better result (as shown in the Right figure).



Listen2Student

Figure: Illustration of our Listen2Student.

(Left) We categorize the regression pseudo-labels into beneficial and misleading instances, and (Right) our Listen2Student prevents the misleading regression pseudo-labels and thus improves the regression branch To understand more details about our method, we encourage the readers to take a closer look at our paper.



Paper



Unbiased Teacher v2: Semi-supervised Object Detection for Anchor-free and Anchor-based Detectors

IEEE / CVF Computer Vision and Pattern Recognition Conference, 2022 (CVPR 2022)
Yen-Cheng Liu, Chih-Yao Ma, Zsolt Kira




Code





Citation

              @inproceedings{liu2022unbiased,
                title={Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors},
                author={Liu, Yen-Cheng and Ma, Chih-Yao and Kira, Zsolt},
                booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
                year={2022}
              }        
            

Acknowledgements

This project is done partially while Yen-Cheng Liu was interning at Meta.
Yen-Cheng Liu and Zsolt Kira were partly supported by DARPA’s Learning with Less
Labels(LwLL) program under agreement HR0011-18-S-0044, as part of their affiliation with Georgia Tech.